In-the-wild attack abuses AI Agent's trust in Markdown

AI security company Forever Security disclosed an in-the-wild attack: attackers planted a malicious AGENTS.md file disguised as engineering notes in the user's Downloads folder; when a coding Agent starts in that directory, it reads the file as project context and is then induced to rewrite its own configuration, allowing the attacker to instruct it to install malware or hand over login tokens. The attack was intercepted on a real customer machine, and Forever's real-time protection blocked the NPM package installation before the malicious code ran; how the file landed was not determined, and researchers say a malicious website may have downloaded it to the user's computer. In testing, the attack chain's success rate exceeded 90% against OpenAI Codex and was about 50% against Claude Code.